Executive brief
CVAT is an open source tool for annotating images and videos used by computer vision teams. A worker-level user could access and cancel automatic annotation requests for tasks they should not have permission to view, and could block other users' annotation work by submitting requests with predictable task IDs. This vulnerability allows unauthorized access to sensitive annotation workflows and enables denial-of-service attacks against other users' tasks.
Technical details
The vulnerability stems from overly lax permission checks in CVAT's lambda request endpoints (retrieve and destroy operations). Request IDs were made predictable through a task-based generation scheme, allowing a worker to enumerate and access requests for any task. The root cause is insufficient authorization validation—viewing and cancelling lambda requests only required the worker role, not access to the underlying task or job. Additionally, a secondary issue allowed users to submit batch lambda requests with inconsistent job and task IDs, enabling denial-of-service by blocking auto-annotation on inaccessible tasks. The fixes implemented in version 2.72.0 restrict viewing requests to users with access to the target task/job and require request cancellation to be performed only by the originating user; a consistency check was added for batch request creation.
Affected products
- CVAT CVAT 2.17.0 to 2.72.0
Timeline
- 2026-08-11: disclosed: CVE-2026-73221 published
- 2026-07-29: patched: Fixed in version 2.72.0